MeshForge / Verification program v8
Twenty-one runnable improvement tracks turn failed image-to-3D candidates into reproducible token, hollow-vessel, sealed-liquid, and coffee-foam benchmarks without granting the pipeline publication authority.
Operating method
The program keeps the original Stable Fast 3D candidate as baseline evidence. Each correction is deterministic, bounded, and scored only after the saved GLB is reopened. Version 8 carries the method through mugs, wine glasses, an original yunomi-inspired milk glass, sealed liquid contents, a separate coffee-foam cap, and a bounded browser surface.
Completed tracks / 01 to 07
Each track owns a specific failure class and leaves evidence that can be inspected independently.
Dataset
Deterministic designs vary bounded body height, circle coverage, recess depth, lip width, and PBR values. Six views of one token remain in the same train, validation, or test split.
Verification
The analyzer identifies graphite body and amber inset roles, then measures height, coverage, recess depth, lip width, bilateral symmetry, and topology after export.
Curriculum
The manifest freezes four stages: hero view, opposed pair, three-view shape set, and full six-view set. Evaluation advances only when the held-out design split improves.
Materials
The canonical GLB assigns separate graphite and amber materials. Each artifact stores base color, metallic, roughness, alpha, transmission, and emission targets for both roles.
Reconstruction
A case-specific repair plan rebuilds the tall hexagonal body, cuts the circular cavity, and positions the amber insert below the top plane while preserving the source candidate.
Shading
The portable contract verifies exported normals and the hard or smooth boundaries encoded by split normals. Logical topology welds coincident attribute splits before counting components and manifold edges.
Material Reconstruction v1
A bounded source sampler records visible color and luminance spread. The analyzer then checks named roles, base color, roughness, metallic response, opacity, transmission, and emission after GLB reimport.
Completed tracks / 08 to 15
These eight tracks add geometry and material contracts that a surface-only silhouette cannot satisfy.
Geometry
The mug and wine-glass generators construct explicit exterior and interior profiles. The verifier measures the opening, rim wall, cavity depth, and bottom after reimport.
Topology
The mug handle remains a separate semantic part with measurable negative space, tube thickness, vertical span, and two body attachment regions.
Profile
A bounded profile generates the bowl, stem, and foot as distinct parts. The analyzer verifies overall height, opening ratio, wall thickness, stem thickness, foot width, and bilateral profile symmetry.
Optics
The exported glTF material carries calibrated transmission, index of refraction, roughness, opacity, metallic response, and emission values. Blender reopens the file and checks each optical field.
Benchmark
Three mug variants and three glass variants remain grouped by design across six shape views and four diagnostic views. Eight explainable gates must pass before a reconstruction can be marked review-ready.
Form
An original hand-sized cup profile adds a gentle outward taper, thick rounded lip, deep cavity, and low foot ring. The analyzer measures each normalized proportion after export.
Optics
The named material roles use a warm base color, partial transmission, soft roughness, full alpha, zero emission, and a glass-range IOR. Reimported values must remain inside the bounded milk-glass range.
Regression
The summary now fails unless ceramic mug, wine glass, and Japanese milk glass are all present with verified inspection digests. Each design keeps six shape views and four diagnostic views.
Completed tracks / 16 to 21
These six tracks preserve portable static geometry while keeping optional browser motion and foam texture outside the evidence artifact. The preview uses a fixed top-quarter inspection view so the coffee ring and white foam stay visible.
Geometry
The generator derives a closed manifold liquid profile from the measured vessel cavity, applies a bounded wall gap, and stores fill, clearance, radius, and target volume.
Verification
Blender reopens the GLB and measures surface height, local cavity radius, top planarity, liquid volume, normal coverage, topology, and semantic roles.
Materials
Coffee, red wine, and milk carry separate base color, roughness, transmission, opacity, emission, and IOR targets that are checked after reimport.
Runtime
A 60 Hz height and velocity surface limits catch-up, impulses, height, and velocity. Unsupported GPUs retain the static vessel without losing the downloadable evidence.
Regression
Three fill variants for each vessel and content pairing must reopen with verified evidence and pass every containment, fill, material, part, shading, and topology gate.
Surface
Coffee adds a separate matte manifold above the liquid. The verifier checks surface-area coverage, thickness, contact gap, cavity containment, rim clearance, material identity, topology, normals, and semantic role after GLB reimport.
Pilot v4 / Reopened evidence
The pilot generated six GLBs, 36 shape views, and 24 material-evidence renders. Blender reopened every GLB and verified two logical manifold components with zero logical non-manifold edges.
Vessel pilot v2 / Reopened evidence
The pilot generated three ceramic mugs, three wine glasses, and three Japanese milk glasses, plus 54 shape views and 36 diagnostic renders. Blender reopened every GLB and verified geometry, hollow volume, handle, optical response, materials, shading, topology, and semantic parts.
Liquid pilot v2 / Reopened evidence
The pilot generated three coffee mugs, three red-wine glasses, and three Japanese milk glasses, plus 54 shape views and 39 diagnostic renders. Blender reopened every GLB and verified containment, fill level, foam coverage and thickness, volume, materials, semantic parts, explicit normals, and topology.
Claim boundary
The material and optical targets are calibrated workflow values, not laboratory measurements recovered from one photograph. Liquid results verify sealed volumes, a bounded foam cap, and an optional preview. They do not establish fluid inference, bubble dynamics, foam decay, or a full simulation. The original neural candidates remain preserved as baseline evidence. High scores still require human review and grant no publication authority.
